Robust Vocal Quality Feature Embeddings for Dysphonic Voice Detection